MTCNN: A Deep Neural Network for Recognizing Stochastic Phase Shifted Time Series Data

Abstract

Time series data is very popular and common in the world, which is widely produced in various real time monitoring systems, such as satellites, power plants, cardiac electrophysiology, and financial transactions. In these applications, it is a critical problem to precisely recognize pattern of a random window on a growing time series stream. Since the data in the window has a stochastic phase shift with the standard known pattern. That leads to a great recognition error for the existing methods. This paper presents a Multi-channel-scale Time series Convolutional Neural Network (MTCNN) to recognize the patterns of time series data with a stochastic phase shift. The MTCNN has several channels with different size convolution kernels so that each layer packs multiple different convolution kernels and pooling structures. It is believed that this deep neural network broadens layers to enhance capability of adapting stochastic phase shift. The experimental results show that the MTCNN is superior to the existing methods, including the Nearest Neighbor method based on Euclidean Distance, the Nearest Neighbor method based on Dynamic Time Warping, Multi-Layer Perception, and the common Convolutional Neural Network.